Justin Pinkney, who works as a software consultant at MathWorks in the UK, has been spending most of his time writing code, developing algorithms, training models, and encouraging people to embrace good software development practices.

Just recently, Justin and a fellow deep learning explorer Doron Adler developed a new “Toonify Yourself” system that got the internet buzzing big time. To find out more about their project, Bored Panda spoke to Justin, who said he has been messing around with GANs for quite a while and has trained lots of different models.

Justin said he came up with the idea of modifying networks “using 'layer swapping' to make realistic Ukiyo-e portraits.” He shared some code for this online and Doron Adler tried it out on some of his models, “showing really cool results of these almost real cartoon faces.”

Since then, they have been working together to train some new neural networks and produced the Toonify website.

Justin said that although people seem to have a lot of fun seeing what the neural network will turn their face into, they often mention that the results aren’t as good as a human artist could produce. “That's not really surprising, I don't think a human artist could produce 25,000 of these an hour (which we were doing at the peak of traffic),” he explained.

He is now working on making the website come back for free, which he promised will happen super soon. “Server running costs were the biggest challenge. These are big neural networks that are running, and although Google Cloud makes it easy to scale up to lots of traffic, it's still not cheap,” Justin explained.

He also put up a donation page, which you can check out right here, for access to the live site in order to help keep the site running.

Justin believes that there are lots of possibilities for what AI and deep learning technologies could do, and he thinks people are only just scratching the surface.

“Especially with so much code being open source and GPU resources being so accessible (via Google Collab), it means creators and artists can really easily try this stuff out.”

According to the deep learning explorer, the next steps for GANs are “in giving people more control over the images generated and making them more efficient in terms of training data and computing resources.”